Face Attribute Recognition Based On Deep Neural Network

In recent years,cities of all sizes have been covered with various types of surveillance cameras,and the massive amount of surveillance images has greatly enriched the application scenarios of the face and pedestrian analysis tasks.In this context,the research of intelligent video understanding technology,which allows computers to automatically identify targets and achieves accurate warning and rapid response to accidents,has gradually become one of the core requirements in the security field.In the field of video surveillance pedestrians are the main surveillance objects.As the important information of face,face attributes are widely used in tasks such as face recognition and face retrieval.Face attribute recognition is defined as given a face image,after feature extraction,a set of appearance attributes are predicted.According to the type of attributes,the attributes can be divided into global attributes and local attributes.For local attributes,local regions should be firstly located and local features should be used to improve the attribute recognition accuracy.In addition,in practical application scenarios,due to the influence of environment and camera quality,the pictures taken by surveillance cameras are mostly low-resolution pictures,and such images lack facial detail features,so conventional face attribute recognition methods are not effective in processing low-resolution pictures.For the above problems,this thesis contains the following:(1)A face attribute recognition network based on end-to-end weakly supervised regional localization is proposed.In this paper,an Attribute Localization Module(ALM)is designed in the network.Different from existing methods,ALM can focus on local regions using only image-level attribute annotations and improve face attribute recognition by utilizing regionbased local features.Moreover,a bottom-up skip connection structure is also introduced to fuse the features of multiple convolutional layers,which can enhance attribute-specific regions localization with local details supplement.(2)To deal with the low-resolution face attribute recognition,a multi-task generative adversarial network(MTGAN)is proposed which can be end-to-end trained.In the MTGAN,the generator is a super-resolution network,which can up-sample low-resolution images into fine-scale ones and recover detailed information.The discriminator is a multi-task network,which describes each super-resolved image with a real/fake score,and predicts face attributes at the same time.Furthermore,to make the generator recover more details for easier prediction,the attribute losses in the discriminator are back-propagated into the generator during training.(3)In this paper,a Face multi-attribute dataset(FMAD)is collected,which contains images taken by surveillance cameras in real scenes.There are 17,756 images in total,and each image is labeled with 13 attributes.To improve the efficiency and accuracy of labeling,we also developed an attribute annotation tool,which is easy to operate,friendly interface,and highly expandable.